Article

Identification of a Retinal Circuit for Recurrent Suppression Using Indirect Electrical Imaging

Highlights Authors d Polyaxonal (PAC) type is electrically coupled to Martin Greschner, ON parasol cells Alexander K. Heitman, Greg D. Field, ..., Alexander Sher, Alan M. Litke, d PAC responses closely mirror those of coupled ON parasol E.J. Chichilnisky cells Correspondence d PAC spikes propagate outward, producing GABA-ergic inhibition of ON parasol cells [email protected] d PAC network produces reciprocal peripheral suppression in In Brief ON parasol population Greschner et al. discovered an amacrine subcircuit that recurrently modulates the activity of retinal ganglion cells in the magnocellular visual pathway of primates. The amacrine cells are electrically coupled to ON parasol cells, and their signals propagate over long distances, producing reciprocal inhibition of distant ON parasol cells.

Greschner et al., 2016, Current Biology 26, 1935–1942 August 8, 2016 ª 2016 Elsevier Ltd. http://dx.doi.org/10.1016/j.cub.2016.05.051 Current Biology Article

Identification of a Retinal Circuit for Recurrent Suppression Using Indirect Electrical Imaging

Martin Greschner,1,2,* Alexander K. Heitman,2 Greg D. Field,2,3 Peter H. Li,2 Daniel Ahn,2 Alexander Sher,4 Alan M. Litke,4 and E.J. Chichilnisky2,5 1Department of Neuroscience, Carl von Ossietzky University, Oldenburg 26129, Germany 2Systems Neurobiology, The Salk Institute for Biological Studies, La Jolla, CA 92037, USA 3Department of Neurobiology, Duke University School of Medicine, Durham, NC 27710, USA 4Santa Cruz Institute for Particle Physics, University of California, Santa Cruz, Santa Cruz, CA 95064, USA 5Departments of Neurosurgery and Ophthalmology and Hansen Experimental Physics Laboratory, Stanford University, Stanford, CA 94305, USA *Correspondence: [email protected] http://dx.doi.org/10.1016/j.cub.2016.05.051

SUMMARY types of retinal ganglion cells (RGCs), activity due to excitation in the receptive field center can be modulated by stimuli far Understanding the function of modulatory inter- outside the receptive field. An interesting candidate for the circuit neuron networks is a major challenge, because mechanism is a collection of anatomically identified polyaxonal such networks typically operate over long spatial amacrine cell (PAC) types, which have radially projecting axonal scales and involve many neurons of different types. arbors that extend for millimeters over the retinal surface, result- Here, we use an indirect electrical imaging method ing in long-range connectivity well suited to serve as a mecha- to reveal the function of a spatially extended, recur- nism for peripheral response modulation. The PACs are thought to pool rectified synaptic inputs from bipolar cells and probably rent retinal circuit composed of two cell types. This provide inhibitory synaptic inputs to RGCs and bipolar cell termi- recurrent circuit produces peripheral response sup- nals [11, 12]. Their long axonal processes probably require that pression of early visual signals in the primate mag- these cells fire spikes to transmit their signals. Recently, we iden- nocellular visual pathway. We identify a type of tified and characterized a spiking PAC type in the primate polyaxonal amacrine cell physiologically via its that could play such a role [13]. However, no interaction with distinctive electrical signature, revealed by electrical the five major primate ganglion cell types was observed. This coupling with ON parasol retinal ganglion cells re- suggests that other amacrine cell types that have not yet been corded using a large-scale multi-electrode array. recorded may play an important modulatory role in the neural Coupling causes the amacrine cells to fire spikes pathways responsible for high-resolution vision. that propagate radially over long distances, produc- Here, we exploit the fact that several amacrine cell types are ing GABA-ergic inhibition of other ON parasol cells re- gap junction coupled to specific RGC types (e.g., [14, 15]) to target and characterize a previously unrecorded PAC in the pri- corded near the amacrine cell axonal projections. We mate retina that is electrically coupled to ON parasol ganglion propose and test a model for the function of this ama- cells [16, 17]. ON parasol cells are frequently sampled in extra- crine cell type, in which the extra-classical receptive cellular recordings [18] and subserve one of the major primate field of ON parasol cells is formed by reciprocal inhi- retino-geniculate visual pathways [19]. Using large-scale multi- bition from other ON parasol cells in the periphery, electrode recordings and a novel analysis of electrical imaging via the electrically coupled amacrine cell network. that leverages coupling, we reveal a type of PAC on the basis of coupling to ON parasol cells. We find that these PACs exhibit electrical features broadly similar to the PACs previously studied INTRODUCTION and closely mirror the activity of the ON parasol cells to which they are coupled. Furthermore, they transmit their spiking output A major feature of is recurrent response signals over several millimeters of the retina, modulating the re- normalization, which accounts for diverse aspects of modulation sponses of other ON parasol cells near their axons via GABA- of light responses outside of the classical receptive field of visual ergic inhibition. This circuit architecture yields a surprising neurons [1, 2]. However, understanding the specific cell types hypothesis for the function of the PACs—namely, that they pro- and circuits that produce such normalization is difficult because vide recurrent inhibition of visual signals in ON parasol cells of the challenges associated with recording many interacting rather than feed-forward inhibition from the bipolar cell network. cells of different types over large regions. We propose and test a quantitative model in which the extra- In the retina, response modulation outside the classical recep- classical receptive field of ON parasol cells is formed by this tive field has long been observed [3–10]. Specifically, in several recurrent inhibitory mechanism.

Current Biology 26, 1935–1942, August 8, 2016 ª 2016 Elsevier Ltd. 1935 Figure 1. The Electrical Image of ON Parasol Cells Reveals Unexpected Radially Propagating Axons (A) Mosaic of simultaneously recorded ON parasol cell receptive fields (top) (scale bar, 200 mm), overlaid spike-triggered average response time course (lower left) and autocorrelation of firing (lower right). (B) Electrical image of a single ON parasol cell (marked dark gray in A), on the multi-electrode array (32 3 16 electrodes, 60-mm spacing). Size of disk and gray value represent minimal voltage deflection on each electrode. Largest amplitudes are saturated to increase visibility of PAC axons. (C) Average voltage traces of electrodes marked in (B) during a recorded spike in the ON parasol cells, with characteristic somatic (trace 1) and axonal waveforms of ON parasol cell (trace 2) and PAC (traces 3–7). Blue and red regions indicate times before and after a time point 0.85 ms after the largest negative voltage deflection at the soma. (D) Electrical image before (blue) and after (red) a time point 0.85 ms after minimal voltage deflection, revealing distinct spatial patterns.

RESULTS observable on the multi-electrode array)—is consistent with only one known cell class in the retina: spiking PACs [13]. These Multi-electrode Recordings and Indirect Electrical auxiliary propagating signals were present in recordings of both Imaging of PACs spontaneous and visually driven activity. To identify amacrine cell types interacting electrically with RGCs, Why would the electrical image of an ON parasol cell reveal the we analyzed the electrical signatures of RGCs in large-scale superimposed electrical signature of a PAC? Spike-sorting arti- multi-electrode recordings. The spikes of hundreds of RGCs of facts were unlikely, given the discrete clusters formed by the several types were sorted on the basis of their characteristic parasol cell spike waveforms and the fact that these additional spatiotemporal waveforms [20, 21] (see the Experimental Proce- PAC spikes were observed only in ON parasol cells and not in dures). ON and OFF parasol cells, two of the five numerically OFF parasol cells or ON or OFF midget cells. The most likely dominant RGC types in primate retina, were identified by their origin of these signatures is electrical coupling to a PAC. Specif- characteristic density and light responses [18, 21]. Then, using ically, because the electrical image is calculated by averaging all the spike times of a given parasol cell, the average voltage voltage waveforms recorded at the times of ON parasol cell waveform over all electrodes near the time of the spike was spikes, voltage fluctuations in any cell that fires spikes in a pre- calculated [13, 20, 22](Figures 1B and 1C). This electrical image cise temporal relationship to the ON parasol cell spikes will revealed the average spatiotemporal pattern of voltage deflec- appear in the electrical image. Previous anatomical work has re- tions recorded over the array during and after the occurrence vealed a type of PAC that is tracer coupled, and thus presumably of spikes in the parasol cell. The averaging procedure made it electrically coupled, to ON parasol cells [16, 17]. Strong electrical possible to detect and characterize small voltage deflections coupling would be expected to produce spikes in these PACs in produced in association with the firing of the cell, including sig- a precise temporal relationship to spikes in the ON parasol cell. nals far from the soma. Based on these considerations, the auxiliary axons in the electri- As expected from previous work [13, 20, 22], the large cal images of ON parasol cells were interpreted as revealing biphasic somatic spike produced by the parasol cell was sur- spikes from electrically coupled PACs. rounded by a smaller opposite-sign dendritic potential on nearby electrodes and was followed by a triphasic axonal spike prop- Physiological Properties of Coupled PACs agating across the array toward the (Figures 1B The interpretation that the auxiliary axons originate from electri- and 1C). These general features were common to all RGCs in cally coupled PACs was supported by the blockade of gap the recording. junction coupling using meclofenamic acid (20 mM) [23, 24](Fig- However, on closer inspection, the electrical images of ON ure 2A). This blockade reduced the magnitude of auxiliary axon parasol cells revealed an additional structure, indicating a strong signals in the electrical image relative to the strength of the interaction with a second cell type. In addition to the main axonal primary axon signal. Since the electrical image is the average spike (Figure 1C, trace 2), smaller auxiliary spikes were detected, voltage waveform around the time of a somatic spike of the refer- emanating from a region near the ON parasol cell soma and ence cell, this can be interpreted as reduced coupling of the PAC propagating simultaneously, and more slowly, in multiple direc- and the ON parasol cell, causing them to fire together less tions (Figure 1C, traces 3–7) (see Movie S1). The electrical image frequently. Gap junctions are commonly found between various of the ON parasol cell and the auxiliary spike axons could be cell types in the retina, and the effects of gap junction blockers separated in time (Figure 1D), revealing the distinctive signature are generally hard to interpret. However, in this case, the electri- of the auxiliary signal. The pattern of radial spike propagation in cal image permitted separate analysis of the coupling of the ON this signal—in some cases, over 3 mm (the maximum distance parasol cell and the PAC, because the same basic features

1936 Current Biology 26, 1935–1942, August 8, 2016 unimodal, whereas unreliable coupling would be expected to produce a distribution composed of voltages recorded with and without PAC spikes. The observed distribution can be modeled using variance from other portions of the voltage trace (Figure 2B). The coupling efficiencies estimated in this way var- ied from cell to cell (mean ± SD: 57% ± 17%; 61 ON parasol cells). Furthermore, in agreement with earlier results from different types of PACs that were directly recorded [13], the spike con- duction velocity of PAC axons was much slower than that of ON parasol cell axons (0.5 ± 0.1 m/s and 1.5 ± 0.2, respectively; 28 cells) (Figure 3A). As in other PAC types, spike conduction ve- locity declined as a function of distance from the soma, unlike in ON parasol cell axons. To analyze the conduction velocity, we Figure 2. Electrical Coupling between ON Parasol RGCs and PACs examined the timing of axon spikes as a function of distance Produces the Radial Axonal Electrical Images (A) Relative (Rel.) estimated coupling strength between ON parasol cells and (Figure 3B). In all cases examined, this analysis revealed a delay PACs (i.e., relative signal amplitude in the electrical image) in the presence and between the ON parasol cell spike and the PAC spike initiation, absence of gap junction blocker meclofenamic acid (MFA, 20 mM). Error suggesting that the parasol cell spike predominantly caused bars indicate SD over seven cells. Values are shown relative to the control the PAC spike. condition. (B) Unimodal distribution of voltage values indicates high coupling efficiency Modulation of Visual Signals in ON Parasol RGCs between PACs and ON parasol cells. Distribution of peak voltage deflection values on a single electrode along the PAC axon are shown in black. Model Many types of amacrine cells, including PACs, are known to prediction for various coupling coefficients while maintaining identical mean exert an inhibitory influence on RGCs [25]. To test whether the amplitude. Best-fitted estimate of the observed distribution is given by a firing of this PAC type modulated the firing of simultaneously re- coupling coefficient (probability of PAC spike given the occurrence of an ON corded RGCs, in recordings of spontaneous activity, we corre- parasol spike) of 86%. lated the spike times of the PAC (inferred from the spike times of the recorded ON parasol cell) with those of other RGCs. appeared in the electrical image but with a different balance. For ON parasol cells, this analysis revealed a spatially specific Furthermore, the reduction of PAC signal strength in the electri- pattern of interaction with the PAC (Figure 4A). On average, the cal image would not be expected from an overall suppression of ON parasol cells lying near the axons of the PAC exhibited firing by the drug but would be expected based on reduced reduced firing after the occurrence of an inferred PAC spike, coupling. A concentration of 100 mM meclofenamic acid led to while parasol cells farther from the axons did not (Figures 4C the complete disappearance of the auxiliary axon signals and and 4D), suggesting that the PAC provided inhibitory syn- could be interpreted as a complete uncoupling of the ON parasol aptic input to other ON parasol cells. This interpretation was cell and PAC; however, at this concentration, the overall supported by blockade with the GABAA receptor antagonist ga- responsiveness and health of the retina decreased strongly, bazine (5 mM) (Figures 4E and 4F). In sum, these findings are and recovery after wash could not be achieved. consistent with the hypothesis that PACs are electrically coupled The number of ON parasol cells that displayed detected PAC to ON parasols near the ON parasol cell soma and form GABA- axons varied across preparations. In some preparations, PAC ergic synapses with other ON parasol cells along their axons axons could be detected in the electrical images of over 90% (Figure 5A) [17]. of ON parasol cells (e.g., in 158 out of 167 ON parasol cells in Fig- How does response suppression by PACs modulate visual ure 4A). The largest PAC axonal spikes identified in the electrical signals transmitted to the brain by the ON parasol cells? This images of ON parasol cells were up to 10% of the amplitude of was examined by analyzing the modulation of ON parasol cell ON parasol axonal spikes. These indirectly recorded PACs firing in the presence of a white-noise visual stimulus. To tease (detected by their electrical coupling to ON parasol cells) were apart the effects of the visual stimulus and PAC inhibition, we observed simultaneously with directly recorded PAC types fitted a generalized linear model (GLM) to the recorded re- described previously [13]. Directly recorded overlapping PACs sponses [26]. In the GLM, stimulus-driven responses sum with exhibited axonal arbors spatially distinct from those in the spike-driven feedback from other cells to produce spikes via parasol-coupled PACs. Furthermore, in some preparations, the a nonlinear mechanism (Figure 5B). The spatiotemporal light receptive fields of the directly recorded PACs formed a mosaic, sensitivity of the GLM, and the sign and time course of the indicating that nearly all cells of this type in the region had been spike-dependent feedback were fitted to data. Sample fits (Fig- recorded and, therefore, that the parasol-coupled PACs were of ure 5C) show that feedback associated with PAC spikes sup- a different type. The indirectly recorded PACs were observed at pressed ON parasol firing in the presence of the stimulus, with scotopic light levels (data not shown) in addition to the photopic a time course similar to suppression observed with no stimulus light levels reported herein. (Figure 4B). The coupling filters of PACs to ON parasol cells lying Examining the distribution of voltage amplitudes at the elec- near their axons were suppressive, on average, across cells, trodes near PAC axons yields a rough estimate of the reliability while those of non-overlapping cells were not (Figure 5D). of electrical coupling. If every ON parasol spike triggered a Thus, response suppression by PACs is likely to modulate visual PAC spike, the voltage distribution at an electrode should be signaling in ON parasol cells.

Current Biology 26, 1935–1942, August 8, 2016 1937 Figure 3. Spike Conduction Velocity of PAC Axons Was Much Slower Than that of the ON Parasol Axon (A) Distribution of conduction velocities of ON parasol cell axons and PAC axons in a single recording (28 cells). Means ± SD:1.5 ± 0.2 m/s and 0.5 ± 0.1 m/s, respectively. (B) Voltage waveforms from electrodes obtained at different distances from the soma along the ON parasol cell and PAC axon. Time of spike initiation at the soma was linearly extrapolated toward zero from axon spike times (ON parasol axon blue line, PAC axon red), resulting in an estimated delay be- tween ON parasol and PAC spike initiation. Den- dritic waveforms were not used and are indicated in light gray. Waveform amplitudes are normalized.

The strong electrical coupling between the identified PACs surround is assembled from the recurrent responses of other and ON parasol cells suggests a surprising potential architecture ON parasol cells rather than a distinct network of interneurons. for visual signal modulation—namely, that PAC spikes predom- A variety of peripheral response suppression effects have inantly convey recurrent visual signals from the surrounding ON been demonstrated in the retina [3–6, 8, 10]. One specific parasol cell population rather than providing feed-forward inhibi- example is the the amacrine cell circuit that signals the motion tion. If this hypothesis is correct, one would predict similar spatial of objects in the visual field relative to motion of the background tuning properties of excitation and PAC-mediated suppression in salamander and rabbit [7, 9]. Although the computation in the ON parasol network. To test this possibility, we stimulated is similar in the circuit studied herein, the circuit architecture is the retina with a central modulating sinusoidal grating pattern different. The computational similarity is that, in both cases, surrounded by a larger modulating background grating [3–9]. the visual features driving the receptive field center and the As expected, the responses of ON parasol cells with receptive features driving surround suppression are the same. This differs fields in the central region exhibited suppression by the surround fundamentally from certain cortical normalization circuits in stimulus. The excitatory tuning of ON parasol cells to sinusoidal which the tuning of the excitatory center and the suppressive gratings was then compared to the tuning of PAC-mediated surround can be very different [27], typically, with surround suppression. ON parasol cells stimulated with sinusoidal grat- tuning being much broader. However, despite the computational ings exhibited an excitatory frequency tuning that peaked at similarity between the two retinal circuits, the circuit architecture 0.38–0.77 cycles per degree (cpd) (Figure 6A; in three retinas). is quite different: in the ON parasol network, surround sup- Introduction of a surround grating suppressed firing of central pression is produced by recurrence via electrical coupling, while ON parasol cells, with maximum response suppression occur- in the object-motion-sensing circuit, the suppressive signal is ring at frequencies similar to that of the peak excitatory tuning created by a parallel computation via direct bipolar inputs to (Figure 6B; 0.38–0.77 cpd, in three retinas). The similarity be- amacrine cells. A possible interpretation of these findings is tween ON parasol excitatory frequency tuning and surround that retinal architecture has evolved to perform similar computa- suppression is consistent with recurrent lateral inhibition medi- tions using two different circuit architectures. Another possibility ated by PACs. To further test the recurrent inhibition hypothesis, is that details of the timing or spatiotemporal tuning of surround we used a re-fitted GLM (see the Experimental Procedures)to suppression differ from those of the center tuning in the object- predict responses to the same combinations of central and motion-sensing circuit in a way that has not yet been fully peripheral modulating gratings. By using a strong representative appreciated. Certainly, in the recurrent network revealed here, PAC coupling filter in the model, we obtained a rough upper- surround suppression is precisely tuned to the same spatiotem- bound estimate of the number of surrounding PACs from which poral features as those of the receptive field center (Figure 6), a given ON parasol cell receives a suppressive signal: 55, 52, and because both signals arise from the same ON parasol cell 16 in three retinas. This modeling approach reproduced the fre- type. However, while the PACs closely mirror the activity of quency tuning of surround suppression (Figure 6C), consistent the ON parasol cells to which they are coupled, they very likely with an origin of the suppressive signal in PACs electrically receive additional inputs from bipolar and other amacrine cell coupled to remote ON parasol cells, i.e., a recurrent inhibitory types as well (e.g., [28]). network architecture. The anatomical identity of the recorded cells is uncertain. Earlier dye coupling studies reported that ON parasol cells DISCUSSION were coupled to two distinct amacrine cell types: (1) a smaller type with thin, sparsely branching dendrites that extend for The indirect electrical imaging method developed here revealed 1–2 mm and (2) a larger axon-bearing type with axons extending the recurrent modulatory function of a network of amacrine cells at least for 2–3 mm [16, 17]. The amacrine cell type reported here in the primate retina. The electrical coupling from ON parasol is, most likely, the large, axon-bearing type. The observed auxil- cells to this amacrine cell type, as well as the consequent iary axons usually extended beyond the border of the recording suppression of firing in other ON parasol cells over a large region, area (maximum 3 mm). No evidence for two distinct popula- implies that a component of the extra-classical receptive field tions of coupled amacrine cells was observed. The mean spike

1938 Current Biology 26, 1935–1942, August 8, 2016 Figure 4. Spontaneous Activity of ON Parasol Cells near PAC Axons Is Suppressed by PAC Firing (A) Electrical image of one PAC (red) coupled to one ON parasol cell (receptive field in blue), aligned with receptive field mosaic of all simultaneously recorded ON parasol cells. Grayscale for ON parasol cells indicates amplitude of negative correlations of spontaneous activity between that cell and the original ON parasol cell (minimum observed correlation coefficient = 0.0042). Correlations were measured in an 8-ms time interval, shifted according to the conduction time along PAC axon. Electrical image is shown for 2–6 ms after the ON parasol somatic spike, as in Figure 1D, lower panel. Scale bar, 1 mm. (B) Cross-correlation function for ON parasol cells overlapping with one PAC axon, revealing deflections suggestive of response suppression. Distance between putative PAC soma and ON parasol cells increases from the top to the bottom panels. Bin size, 1 ms. (C and D) ON parasol cells with somas near and far from the axons of the PAC were suppressed and not, respectively. Near and far ON parasol cells were separated based on their electrical image overlap with the electrical image of the PAC. Traces show cross-correlation functions, averaged across cells after adjustment for conduction time along the PAC axon. Neighboring ON parasol cells exhibited strong positive correlations due to shared input and direct coupling (see the Experimental Procedures) and, thus, were excluded from analysis. Bin size, 1 ms; shaded area, ±1 SD, 18 overlapping, 118 non-overlapping cells. (E) As in (C), different preparation. Shaded area, ±1 SD, ten ON parasol cells and one PAC. (F) Negative correlations blocked by 5 mM gabazine (SR-95531). Same as in (E). Separate preparations were used for (A), (B)–(D), and (E) and (F). waveform and spike propagation speed closely resembled those used here. However, the most prominent directly recorded PAC of amacrine cell types described earlier [13], which have a poly- type in recordings performed using the same methods has an axonal structure and probably produce classical sodium spikes OFF light response type [13], indicating that amacrine cells with [29–31]. the OFF response polarity are recordable with this approach. The degree of observed coupling was variable, as in many Furthermore, previous work revealed homotypical coupling in experimental studies involving gap junctions. This variation ON parasol cells, but not OFF parasol cells [32, 33]. could reflect differences in the adaptive or circadian state of The recurrent inhibition observed here seems unlikely to the retina. Earlier studies indicated that a parasol cell is dye mediate visual functions that rely on precise spike timing. Every coupled to multiple PACs [16, 17]. In the present data, the elec- ON parasol cell was inhibited by the low-pass filtered mean ac- trical images of ON parasol cells were usually dominated by one tivity of roughly 50 distant ON parasol cells. Over the >3-mm or two cells that were presumably strongly coupled (see Figures length of PAC axons, spike propagation time can be >6 ms (Fig- 2A and 4A). Variations of coupling strength within a preparation ure 3A), which could introduce large temporal offsets between could reflect a range of distances between the cell bodies of different inhibitory PAC inputs to a given ON parasol cell. Also, ON parasol cells and the nearest PACs of the coupled popula- every PAC spike produced inhibition lasting roughly 15 ms. tion, which could influence the probability and magnitude of Therefore, it seems that this mechanism produces inhibition gap junction coupling. reflecting aggregate statistics of responses, and may not be It remains to be seen whether the recurrent architecture is appropriate for precisely timed inhibitory functions. specific to ON parasol cells or more general in the retina. In the Recurrent normalization of responses similar to the present present data, among the major cell types (ON midgets, OFF results is thought to be important for diverse neural computa- midgets, ON parasol, OFF parasol, and small bistratified cells), tions in [1, 2]. The present findings indicate that the ON parasol cells were the only cell type that showed electrical this important neural computation begins in the retina. coupling to an amacrine cell. However, an earlier report on the tracer coupling between parasol and amacrine cells made no EXPERIMENTAL PROCEDURES distinction between the coupling pattern of ON and OFF parasol cells [16], leaving open the possibility of a similar network for OFF Retinas were obtained and recorded as described previously [20, 21]. Briefly, parasol cells. It is possible that existing PAC axons in the OFF were taken from terminally anesthetized macaque monkeys (Macaca layers were not deleted with the extracellular electrode arrays mulatta, M. fascicularis), used by other laboratories in the course of their

Current Biology 26, 1935–1942, August 8, 2016 1939 Figure 5. The PAC Network Recurrently Inhibits Light Responses of ON Parasol Cells (A) Schematic model of hypothesized PAC-mediated lateral inhibitory network. (B) Schematic of the generalized linear model (GLM) used to summarize and analyze light responses and lateral inhibition of ON parasol cells. The spike rate controlling the model output is a nonlinear function of the sum of the filtered stimulus, filtered spike trains representing inputs from PACs, and a post-spike feedback filter internal to the cell. (C) Individual fitted coupling filters representing PAC inputs, which modulate the gain of light-driven inputs to the modeled ON parasol cell. All filters were fitted to data obtained in the presence of a white-noise stimulus. (D) Average coupling filter associated with 20 ON parasol cells overlying PAC axon (dark line) and mean ± SD of filters associated with 20 ON parasol cells (resampled from 150 cells) not overlying PAC axon (gray range). Filters adjusted for conduction time along the PAC axon prior to averaging. The same preparation as in Figures 4B–4D was used. experiments, in accordance with institutional guidelines for the care and use of cally distinct ganglion cell types (e.g., ON parasol) were inferred based on animals. Segments of peripheral retina were isolated from the pigment epithe- cell densities and light response properties [18]. lium and placed flat, ganglion cell side down, on a planar array of extracellular Electrical images were calculated as the spike-triggered average electrical microelectrodes. Two different electrode arrays of 512 electrodes were used. signal across the array for each cell separately [13, 20, 22]. Specifically, for One array covered a rectangular region 1,890 3 900 mm, and the second each electrode, the voltage waveform during the time period from 0.5 ms covered a hexagonal region 3,120 mm wide. While recording, the retina was before to several milliseconds after each spike from the cell was identified, perfused with Ames’ solution (33–36 C) bubbled with 95% O2 and 5% CO2 and the average across spikes was computed. For display of the electrical im- (pH 7.4). The light intensity was kept constant across experiments at 830 age, the minimum voltage deflection at each electrode was shown using the (840, 440) photo-isomerizations (L [long] (M, S [medium, short]) cone)1 s1. spatial layout of the electrodes, represented by the size of disk and its gray The display was calibrated using a PR-701 spectroradiometer (Photo- value. Largest amplitudes were saturated to increase visibility of PAC axons. Research) and a photodiode (UDT Instruments). The photopic photon absorp- The conduction velocity of action potentials in ON parasol cells and PAC axons tion rates were computed assuming a 0.37-mm2 collecting area for primate (Figure 3) was estimated by a linear fit to the distance the spike traveled along cones [34]. In addition, the indirectly recorded PACs were observed across the axon per unit time. light levels from 4,500 photo-isomerizations rod1 s1 (corresponds to 1,550 To evaluate the effects of the gap junction blocker meclofenamic acid (1,570, 740) photo-isomerizations (L (M, S) cone)1 s1) to 0.05 photo-isomer- (Sigma; Figure 2A), we measured changes in amplitude of the electrical image izations rod1 s1 (data not shown). of axonal spikes between control, drug, and washout conditions. The electrical Recordings were analyzed offline to isolate the spikes of different cells, as image is the average voltage waveform around the time of a somatic spike of described previously [18, 21]. Briefly, candidate spike events were detected the reference cell. Thus, any reduction in the synchronized firing of PAC and using a threshold on each electrode, and the voltage waveforms on the elec- ON parasol cells due to loss of coupling would be expected to reduce the trode and neighboring electrodes around the time of the spike were extracted. contribution of the PAC to the ON parasol electrical image. Therefore, the Clusters of similar spike waveforms were identified as candidate neurons reduction in the magnitude of apparent PAC axons was interpreted as a reduc- if they exhibited a refractory period. Duplicate recordings of the same cell tion in coupling. Meclofenamic acid was used in a low concentration (20 mM) to were identified by temporal cross-correlation and removed. enable a washout. At higher concentrations (100 mM), PAC axons were no A white-noise stimulus, composed of a lattice of square pixels updating longer detected, but the stability of the recording was negatively impacted randomly and independently of one another over time, was used to charac- [36]. So that a sufficient number of spikes could be recorded for the subsequent terize the spatial, temporal, and chromatic response properties of recorded analysis, the drug was administered for 30–60 min. During this time, the overall cells [35]. The spike-triggered average stimulus obtained in response to white spike rate decreased, particularly in ON cells. At high concentrations, no noise was taken as a summary of the spatiotemporal-chromatic light response reversal of the observed effects could be obtained with a washout of over 1 hr. properties of the cell. The receptive fields were summarized by fitting the Cross-correlation functions (Figures 4B–4F) were obtained by binning spike-triggered average with a two-dimensional elliptical Gaussian function. spikes into 1-ms time bins and computing the correlation coefficient between These receptive fields fits were visualized by an ellipse for each cell represent- the resulting spike count vectors as a function of temporal offset. ON parasol ing the 1 SD contour of the Gaussian (Figure 1A). Correspondences between cells were classified manually into groups that did, and did not, overlap with a functionally defined ganglion cell types and previously identified morphologi- PAC axon. ON parasol cells very near to the reference cell body were omitted

1940 Current Biology 26, 1935–1942, August 8, 2016 Figure 6. Model of Recurrent Inhibition Re- produces Extra Classical Ef- fects in ON Parasol Cells (A) Spatial frequency tuning of ON parasol cells to full-field sinusoidal drifting gratings (at 4 Hz; n = 89), shaded region ± 1 SD. cyc/deg, cycles per degree. Inf indicates constant gray back- ground. (B) Normalized response to central grating pattern as a function of spatial frequency of a surrounding grating pattern (10 center diameter). Each gray line represents a specific central grating frequency (0.77, 0.38, 0.19, 0.1, 0.05, and 0.02 cycles per degree; at 4 Hz); black line represents average (n = 13 ON parasol cells). (C) Refitted GLM simulations of surround suppression using fits of the model to data from the same cells as in (B) (see the Experimental Procedures). Insets at the top of each panel indicate the stimulus. because they exhibited strong positive correlations from shared photore- Research Fellowship 1003198 (P.H.L.). We also thank C.K. Hulse for technical ceptor noise and electrical coupling [33, 37]. assistance; M.I. Grivich, D. Petrusca, A. Grillo, P. Grybos, P. Hottowy, and S. To estimate how response suppression by PACs modulates visual signals, Kachiguine for technical development; H. Fox, M. Taffe, E. Callaway, and K. we fitted a GLM to the data (Figure 5B). The GLM consists of a bias term m, Osborn for providing access to retinas; and S. Barry for machining. stimulus filter k, a spike-history filter h, and a set of incoming coupling filters

{li}. The parameters act on the stimulus x, the spike train of the fitted Received: April 7, 2016 neuron y, and the spike trains of other RGCs that could affect that firing rate Revised: May 18, 2016 of the fitted neuron {yi}, producing a spike rate l(t) given by Accepted: May 19, 2016 l = m + + + : Published: July 7, 2016 ðtÞ expð k xðtÞ h yðtÞ fligfyigÞ

The convex nature of the maximum likelihood estimator for the GLM guaran- REFERENCES tees that the optimal model fit is found [26]. A GLM (m, k, h, l) was fitted to ON parasol cells one PAC at a time in the pres- 1. Heeger, D.J. (1992). Normalization of cell responses in cat striate cortex. ence of a white-noise stimulus. Strong coupling filters {li} were found in over 30 Vis. Neurosci. 9, 181–197. cells across three different retinas (Figure 5C). The significant difference be- 2. Carandini, M., Heeger, D.J., and Movshon, J.A. (1997). Linearity and tween the average coupling filter of cells that did and did not intersect PAC normalization in simple cells of the macaque primary visual cortex. axons (Figure 5D) was verified for five PACs across two different retinas. J. Neurosci. 17, 8621–8644. Center-surround gratings were presented to probe the function of the proposed inhibitory PAC circuit (Figure 6B). Six different spatial frequencies 3. McIlwain, J.T. (1964). Receptive fields of optic tract axons and lateral genic- 27 were used in the center, and eight different spatial frequencies and a uniform ulate cells: peripheral extent and barbiturate sensitivity. J. Neurophysiol. , gray were used in the surround. Each combination was shown for 10–12 s in a 1154–1173. randomized order and repeated two to three times. Each center suppression 4. Levick, W.R., Oyster, C.W., and Davis, D.L. (1965). Evidence that profile (gray line) was normalized by the center spike rate, while a uniform McIlwain’s periphery effect is not a stray light artifact. J. Neurophysiol. gray was presented in the surround. The black average line is a standard 28, 555–559. uniformly weighted average of the center suppression profiles. 5. Werblin, F.S. (1972). Lateral interactions at of verte- A GLM (m, k, h) was fit to the white-noise stimulus for all cells that were used in brate retina: antagonistic responses to change. Science 175, 1008–1010. Figure 6B. A strong representative PAC coupling filter, l , with a gain reduction rep 6. Passaglia, C.L., Enroth-Cugell, C., and Troy, J.B. (2001). Effects of remote of 15% was used to simulate PAC inhibition. Each cell then had two free scaling stimulation on the mean firing rate of cat retinal ganglion cells. J. Neurosci. parameters c m; c k; h; l to refit the GLM to the drifting grating stimulus. ð 1 2 repÞ 21, 5794–5803. The parameters c1 and c2 were fit to minimize the squared difference between the observed frequency tuning curve and the GLM simulated tuning curve. Us- 7. Olveczky, B.P., Baccus, S.A., and Meister, M. (2003). Segregation of ob- 423 ing the refitted GLM, we then simulated all of the center cell models with the ject and background motion in the retina. Nature , 401–408. same stimulus conditions and durations that were used in the experiment. Sur- 8. Zaghloul, K.A., Manookin, M.B., Borghuis, B.G., Boahen, K., and Demb, round inhibition was modeled as a Poisson process, with frequency determined J.B. (2007). Functional circuitry for peripheral suppression in mammalian by the measured firing rate of surround cells (Figure 6A) multiplied by the simu- Y-type retinal ganglion cells. J. Neurophysiol. 97, 4327–4340. lated number of PACs assumed to converge onto the ON parasol cell. 9. Baccus, S.A., Olveczky, B.P., Manu, M., and Meister, M. (2008). A retinal circuit that computes object motion. J. Neurosci. 28, 6807–6817. SUPPLEMENTAL INFORMATION 10. Hoggarth, A., McLaughlin, A.J., Ronellenfitch, K., Trenholm, S., Vasandani, R., Sethuramanujam, S., Schwab, D., Briggman, K.L., and Supplemental Information includes one movie and can be found with this Awatramani, G.B. (2015). Specific wiring of distinct amacrine cells in the article online at http://dx.doi.org/10.1016/j.cub.2016.05.051. directionally selective retinal circuit permits independent coding of direc- tion and size. Neuron 86, 276–291. AUTHOR CONTRIBUTIONS 11. Demb, J.B., Haarsma, L., Freed, M.A., and Sterling, P. (1999). Functional circuitry of the ’s nonlinear receptive field. J. Neurosci. M.G., A.K.H., G.D.F., P.H.L., D.A., A.S. and E.J.C. conducted the experiments, 19, 9756–9767. M.G., A.M.L., and E.J.C. designed the experiments. M.G., A.K.H. and E.J.C. wrote the paper. 12. Demb, J.B., Zaghloul, K., Haarsma, L., and Sterling, P. (2001). Bipolar cells contribute to nonlinear spatial summation in the brisk-transient (Y) gan- 21 ACKNOWLEDGMENTS glion cell in mammalian retina. J. Neurosci. , 7447–7454. 13. Greschner, M., Field, G.D., Li, P.H., Schiff, M.L., Gauthier, J.L., Ahn, D., This work was supported by NIH grant EY017992 (E.J.C.), VolkswagenStiftung Sher, A., Litke, A.M., and Chichilnisky, E.J. (2014). A polyaxonal amacrine (M.G.), the Helen Hay Whitney Foundation (G.D.F.), and NSF Postdoctoral cell population in the primate retina. J. Neurosci. 34, 3597–3606.

Current Biology 26, 1935–1942, August 8, 2016 1941 14. Marc, R.E., and Jones, B.W. (2002). Molecular phenotyping of retinal gan- 26. Pillow, J.W., Shlens, J., Paninski, L., Sher, A., Litke, A.M., Chichilnisky, glion cells. J. Neurosci. 22, 413–427. E.J., and Simoncelli, E.P. (2008). Spatio-temporal correlations and visual 15. Vo¨ lgyi, B., Chheda, S., and Bloomfield, S.A. (2009). Tracer coupling pat- signalling in a complete neuronal population. Nature 454, 995–999. terns of the ganglion cell subtypes in the mouse retina. J. Comp. Neurol. 27. Carandini, M., and Heeger, D.J. (2011). Normalization as a canonical neu- 512 , 664–687. ral computation. Nat. Rev. Neurosci. 13, 51–62. 16. Dacey, D.M., and Brace, S. (1992). A coupled network for parasol but not 28. Murphy-Baum, B.L., and Taylor, W.R. (2015). The synaptic and morpho- midget ganglion cells in the primate retina. Vis. Neurosci. 9, 279–290. logical basis of orientation selectivity in a polyaxonal amacrine cell of the 17. Jacoby, R., Stafford, D., Kouyama, N., and Marshak, D. (1996). Synaptic rabbit retina. J. Neurosci. 35, 13336–13350. inputs to ON parasol ganglion cells in the primate retina. J. Neurosci. 16, 8041–8056. 29. Stafford, D.K., and Dacey, D.M. (1997). Physiology of the A1 amacrine: a spiking, axon-bearing interneuron of the macaque monkey retina. Vis. 18. Field, G.D., Sher, A., Gauthier, J.L., Greschner, M., Shlens, J., Litke, A.M., Neurosci. 14, 507–522. and Chichilnisky, E.J. (2007). Spatial properties and functional organiza- tion of small bistratified ganglion cells in primate retina. J. Neurosci. 27, 30. Davenport, C.M., Detwiler, P.B., and Dacey, D.M. (2007). Functional polar- 13261–13272. ity of dendrites and axons of primate A1 amacrine cells. Vis. Neurosci. 24, 19. Perry, V.H., Oehler, R., and Cowey, A. (1984). Retinal ganglion cells that 449–457. project to the dorsal lateral geniculate nucleus in the macaque monkey. 31. Manookin, M.B., Puller, C., Rieke, F., Neitz, J., and Neitz, M. (2015). 12 Neuroscience , 1101–1123. Distinctive receptive field and physiological properties of a wide-field ama- 20. Litke, A.M., Bezayiff, N., Chichilnisky, E.J., Cunningham, W., Dabrowski, crine cell in the macaque monkey retina. J. Neurophysiol. 114, 1606–1616. W., Grillo, A.A., Grivich, M., Grybos, P., Hottowy, P., and Kachiguine, S. (2004). What does the tell the brain? Development of a system for 32. Shlens, J., Rieke, F., and Chichilnisky, E. (2008). Synchronized firing in the 18 the large-scale recording of retinal output activity. IEEE Trans. Nucl. Sci. retina. Curr. Opin. Neurobiol. , 396–402. 51, 1434–1440. 33. Trong, P.K., and Rieke, F. (2008). Origin of correlated activity between 21. Frechette, E.S., Sher, A., Grivich, M.I., Petrusca, D., Litke, A.M., and parasol retinal ganglion cells. Nat. Neurosci. 11, 1343–1351. Chichilnisky, E.J. (2005). Fidelity of the ensemble code for visual motion 34. Schnapf, J.L., Nunn, B.J., Meister, M., and Baylor, D.A. (1990). Visual in primate retina. J. Neurophysiol. 94, 119–135. transduction in cones of the monkey Macaca fascicularis. J. Physiol. 22. Petrusca, D., Grivich, M.I., Sher, A., Field, G.D., Gauthier, J.L., Greschner, 427, 681–713. M., Shlens, J., Chichilnisky, E.J., and Litke, A.M. (2007). Identification and characterization of a Y-like primate retinal ganglion cell type. J. Neurosci. 35. Chichilnisky, E.J. (2001). A simple white noise analysis of neuronal light re- 12 27, 11019–11027. sponses. Network , 199–213. 23. Pan, F., Mills, S.L., and Massey, S.C. (2007). Screening of gap junction an- 36. Kuo, S.P., Schwartz, G.W., and Rieke, F. (2016). Nonlinear spatiotemporal tagonists on dye coupling in the rabbit retina. Vis. Neurosci. 24, 609–618. integration by electrical and chemical synapses in the retina. Neuron 90, 24. Veruki, M.L., and Hartveit, E.(2009). Meclofenamicacidblocks electricalsyn- 320–332. apses of retinal AII amacrine and on-cone bipolar cells. J. Neurophysiol. 101, 37. Greschner, M., Shlens, J., Bakolitsa, C., Field, G.D., Gauthier, J.L., 2339–2347. Jepson, L.H., Sher, A., Litke, A.M., and Chichilnisky, E.J. (2011). 25. Masland, R.H. (2012). The neuronal organization of the retina. Neuron 76, Correlated firing among major ganglion cell types in primate retina. 266–280. J. Physiol. 589, 75–86.

1942 Current Biology 26, 1935–1942, August 8, 2016